ArchitectureA supervised JVM-class runtime — OLTP on seven engines, OLAP on three. AI-native, MCP-native, observable as plain SQL.Read the architecture
Está viendo la edición Perú. Está viendo la edición Colombia. You're viewing the Pakistan edition. Cambiar a la edición global →Cambiar a la edición global →Switch to the global edition →

The MCP surface redrawn: categories as decisions, risk as ceiling, verbs first

Grant categories now name a decision an administrator makes, risk class bounds what a category grant gives, every tool title is a verb and an object, and descriptions got shorter under a lowered token ceiling.

An MCP surface is read by two audiences: the administrator who grants it and the model that calls it. Both read it in pieces, and both are misled by inconsistency. This release redraws the surface so that a category is something a person decides and a title is something a model can predict.

Categories are decisions

  • One per family Categories are named by the gate's area and the family's word. Read and write are no longer separate categories: every tool carries a risk class, and the read half of a family is that family at R0. Fifty-four categories became 48.
  • Too small to decide Five categories held a single tool and seven held two. They merged under the decision they belong to, such as the essentials every developer works with and the deployment of table models and routines into a data database. Forty-eight became 41.
  • Unknown codes refused Grant, revoke and list resolve a category from code and refuse an unknown one, listing the valid ones.

Risk as the ceiling of a grant

  • max_risk on grant A pattern or category grant takes a ceiling from R0 to R3: R0 gives a family's reads, R1 its reads and ordinary writes without deletes. A tool whose prerequisite sits above the ceiling is left out with it, and the plan lists both.
  • min_risk on revoke The mirror: a revoke at R1 removes a category's writes and keeps its reads. A tool named exactly remains the administrator's explicit choice, whatever its risk.

What a client lists

  • Titles Every title follows one rule: sentence case, a verb and an object, the verb matching the tool name's ending, at most 40 characters, each used once.
  • Descriptions Every description opens with its verb, and the paragraph shared by the write and edit tools is said once in fewer words; the dictionary search tool's contract went from 2,617 characters to 1,312.
  • A lowered ceiling The token budget for listing the surface is a ratchet. With the shorter descriptions it came down to 66,200 for development and 30,100 for administration.
  • Instructions in order The connect-time instructions now open with who is connected and the session's scope, then one clause per namespace, because a client may keep only their beginning.

None of this adds a capability. It makes the existing ones grantable by reading the category list and callable by reading the title.